How to Build a Smart Agent on RAXION (Phase 0 Developer Guide)
A technical developer guide to the current PoC stack, the CoherenceScore pipeline, the architecture wrappers, and the upcoming runtime work in Q2.
Phase 0 is still a Python PoC plus a Rust proof scaffold. There is no blockchain yet, no token mechanics in production, and no on-chain program layer.
What exists today is the poc directory for architectures, convergence, benchmarks, and the main runner, plus the proofs directory for the RISC Zero guest and host.
What does not exist yet is the Agave runtime fork, the Anchor programs, the Rust Smart Agent SDK, and the Jolt proof circuits.
The pipeline is five steps: embed each model output, compute pairwise cosine similarity, compute CS_semantic as a geometric mean, approximate CC, and combine both into the final CoherenceScore using the 0.4 and 0.6 weights.
With only two architectures in Phase 0, the geometric mean effectively collapses to the shared pairwise similarity. That limitation is acceptable now, but disappears only when the third architecture is added.
The current CC still uses a neutral prior and a sentence-level approximation, so developers should treat it as a scaffolding layer rather than the final form of the protocol.
The BaseArchitecture interface defines the contract for adding a new architecture to the system: an architecture_type and an infer() method that returns an InferenceOutput.
Today, the PoC runs via Ollama with a Transformer wrapper and an SSM proxy wrapper. The SSM proxy is only a behavioral approximation and is intended to be replaced later with a true Mamba-like architecture.
- βTransformer: llama3.1:8b
- βSSM proxy: mistral:7b
- βNeuro-Symbolic: still an open issue
On the Rust side, developers need Rust 1.75+ and cargo-risczero. The host program feeds private inputs into the guest, and the guest commits the joint hash inside the zkVM.
The IMAGE_ID is the guardrail that proves the verifier checked the right guest program binary, not just any receipt-shaped object.
The guide explicitly calls out good first issues: the Neuro-Symbolic architecture stub, the 100-query MMLU benchmark, the Rust mirror of CoherenceScore, and the RISC Zero latency benchmark across inputs and hardware.
It also makes clear that ALPHA, BETA, and the three thresholds are protocol invariants. If a contributor changes them, tests should fail and force a whitepaper update first.
Q2 moves the project into the Agave fork. That introduces CognitiveAccount types, the Cognitive Scheduler, and the first real runtime-level engineering risk in the project.
The best way to contribute to that phase is to understand the current PoC deeply, then move into runtime discussions with a concrete reading of where those new account types and scheduler hooks need to live.